1108 results about "Neural network architecture" patented technology

The invention relates to power distribution network fault processing equipment applied to the technical field of power distribution, and aims to improve the fault prediction, diagnosis and processing capabilities of a power system. The equipment comprises a pre-training module, a neural network architecture module, a data acquisition layer, an edge calculation layer, a core processing layer, a context learning module, a reasoning processing module, an output and interaction layer and a continuous learning and improvement layer, and is characterized in that a large-scale pre-trained Transform model is utilized, and professional knowledge of a power system is combined; according to the method, deep understanding and analysis of a complex power grid state are realized, through multi-source data fusion, edge calculation and central cooperative processing, the equipment can monitor the power grid state in real time, accurately predict potential faults and provide intelligent diagnosis and repair suggestions, and model pre-training and optimization are performed by adopting self-supervised learning and federated learning technologies; using reinforcement learning to optimize a fault response strategy; the interpretability of the model is realized, and the decision transparency is improved.

An online system trains a transformer architecture by an initialization method which allows the transformer architecture to be trained without normalization layers of learning rate warmup, resulting in significant improvements in computational efficiency for transformer architectures. Specifically, an attention block included in an encoder or a decoder of the transformer architecture generates the set of attention representations by applying a key matrix to the input key, a query matrix to the input query, a value matrix to the input value to generate an output, and applying an output matrix to the output to generate the set of attention representations. The initialization method may be performed by scaling the parameters of the value matrix and the output matrix with a factor that is inverse to a number of the set of encoders or a number of the set of decoders.

The invention belongs to the technical field of image classification neural networks, and discloses a lightweight image classification neural network architecture system based on knowledge distillation. The invention provides a lightweight image classification neural network architecture system based on knowledge distillation. The lightweight image classification neural network architecture system comprises a data set preprocessing module, a teacher model preprocessing module, a subnet model construction module, a personalized training module, a real loss calculation module, a distillation loss calculation module, a total loss calculation module and a student model construction module. The invention provides a lightweight neural network architecture system based on knowledge distillation, and designs a brand new training mode by using the superiority of knowledge distillation for teacher model weight extraction: the knowledge distillation of teacher and student models is combined with a deep compression method; the system for constructing the lightweight network can be used for training various complex image classification neural networks so as to improve the network performance.

The invention discloses a method for inverting key parameters in an underground water pollutant biodegradation numerical model based on a neural network algorithm. Carrying out preprocessing and refined dimension reduction on the data, and processing high-dimensional multi-scale data; building a multilayer neural network architecture under a physical information neural network framework, embedding a related reaction equation of a pollutant natural attenuation process, and coupling a biochemical reaction mechanism and a data driving model; a neural network model is trained, errors in the training process are calculated, and the accuracy of an inversion result is improved by using a sub-when optimization algorithm; integrating observation data and a model prediction result, determining final parameter setting of a numerical model, and carrying out underground water damage real object quantification; comparative verification of a neural network parameter inversion model is carried out, and feasibility and accuracy of the parameter inversion method are ensured. According to the method, the problem that the key parameters of the biodegradation numerical model are difficult to accurately invert in the real object quantification process of underground water pollution under complex geological conditions is solved.

The invention relates to the technical field of slope protection, in particular to a slope protection intelligent detection system based on deep learning. According to the technical scheme, the system comprises a multi-source heterogeneous data sensing module, a data fusion and feature extraction module, a slope state intelligent diagnosis and early warning module, an edge-cloud collaborative computing architecture and a system optimization module. Registration and feature complementation of multi-source heterogeneous data are realized through a multi-modal detection network, an overfitting phenomenon is effectively inhibited through a physical information neural network architecture, risk quantitative evaluation is realized through construction of a dynamic risk evaluation model, early warning response time is shortened in cooperation with a four-level early warning strategy, the false alarm rate is reduced, and the early warning efficiency is improved. Besides, the detection precision of the system in an extreme scene is improved through a physical constraint adversarial training method, so that the environmental adaptability of the system is improved, continuous updating and evolution of the model are realized through an online incremental learning module, and the problem of performance degradation of a traditional system caused by change of geological conditions is solved.

The invention relates to a supporting structure stress state monitoring method based on artificial intelligence, and belongs to the technical field of artificial intelligence and data processing. The method comprises the following steps: acquiring and marking strain data of a supporting structure; after abnormal values are removed, normalizing the multi-sensor data to generate a normalized strain sequence; a state monitoring model is constructed, a deep time sequence neural network architecture is adopted, and the state monitoring model comprises an input layer, a self-adaptive wavelet attention feature mapping layer, a time domain gating convolution module, a global maximum pooling layer, a dynamic feature importance reweighting layer and a full-connection classification layer; inputting a normalized data training model; optimizing a loss function through a quantile interval adaptive learning rate and a momentum updating strategy; after real-time monitoring data is processed, inputting the data into the training model according to time window slices, outputting four types of probabilities, and taking the maximum value as a prediction state; and if a plurality of continuous windows are early-warning and dangerous, triggering the terminal to give an alarm. The accuracy of monitoring the stress state of the supporting structure can be improved.

Certain examples described herein relate to a mapping system. An example mapping system has a differentiable mapping engine to receive image data comprising a sequence of images captured using one or more camera devices of an object as it navigates an environment and a neural network scene representation comprising a neural network architecture trained to map input coordinate tensors indicating at least a point location in three-dimensional space to scene feature tensors having a dimensionality greater than the input tensors. The neural network scene representation is communicatively coupled to the differentiable mapping engine and the differentiable mapping engine is configured to use the neural network scene representation as a mapping of the environment during operation of the differentiable mapping engine.

Generally, the present disclosure is directed to efficient neural network architectures, techniques for constructing new efficient neural networks, and approaches to low-latency execution of neural networks. In an example aspect, the present disclosure provides a more powerful configuration of an inverted bottleneck block that maintains an efficient execution profile. An example search system can parameterize a universal inverted bottleneck block with a first parameter that categorically activates a spatial mixing operation in the unexpanded state. In this manner, for instance, a large variety of different network architectures can be explored using a relatively compact search space that admits high levels of parameter sharing. Further, the present disclosure introduces an efficiency-optimized multi-query attention block.

This disclosure relates to an asset evaluation system that collects multi-modal asset data, extracts asset features using a neural network architecture, and generates personalized value scores based on user profiles. The system analyzes image and textual content to extract asset features, correlates those features with personalization data in user profiles, and generates tailored asset analysis results. The system can predict a value score for each user profile-asset pair in a manner that goes beyond mere consideration of asset pricing, and which accounts for specific personalization parameters stored in each user profile, such as parameters corresponding to the user's technical capabilities and certifications, risk tolerance, investment preferences, timeline requirements, available resources, and potential synergies with existing assets. Additionally, in certain embodiments, the asset evaluation system may employ a distributed architecture that utilizes multiple interconnected processing nodes to efficiently handle large-scale data processing.

AI neural networks are integrated with distributed ledger technology (DLT), including the integration of decentralized oracles, to enhance the functionality, security, and efficiency of systems. The integration incorporates several components and methods comprising novel neural network architectures, data preprocessing techniques, DLT integration approaches, privacy and security measures, consensus mechanisms, and others. The invention enables the development of neural smart contracts, neural consensus algorithms, dynamic neural network training, neural network optimization techniques, hybrid DLT architectures, neural network interpretability methods, and reinforcement learning for neural smart contract execution. By integrating decentralized oracles into the system, the invention has the capability to further enhance the reliability, accuracy, and diversity of data used within the DLT, improving transaction validation, consensus mechanisms, and overall system performance.

A neural network model training method includes: constructing a first neural network architecture, where the first neural network architecture includes M basic unit layers, each of the M basic unit layers includes a plurality of basic units, and the plurality of basic units includes at least a first-type basic unit and a second-type basic unit; and obtaining a target model through training based on datasets respectively corresponding to a plurality of tasks and the first neural network architecture, where the target model includes a plurality of task paths, at least some of the plurality of task paths include N basic units selected from some of the M basic unit layers, and N<M.

The invention discloses a cable life dynamic evaluation system based on multi-physics field coupling, and particularly relates to the field of industrial automation and control systems, which comprises a multi-physics field sensing module, a coupling analysis engine module, a dynamic life evaluation module, a digital twin interaction module and an environmental interference suppression module, through a distributed optical fiber temperature sensor, a capacitive electric field sensor and a magnetostrictive stress sensor, temperature, electric field, magnetic field and mechanical stress data of a cable are collected in real time, multi-physical field characteristics and a cable defect database are matched in real time by using a cross-scale dynamic association algorithm, a damage state is evaluated, and the cable defect detection accuracy is improved. A time sequence neural network architecture is adopted to predict the remaining life, model self-correction is achieved through digital twin comparison, interference is suppressed in combination with an environment-physical field coupling compensation matrix, sensing and evaluation of the health state of the cable, life prediction and continuous optimization of the model are achieved, and the efficiency of cable life evaluation is improved.

Neural network architectures for feature extraction from visual input. In at least one embodiment, a neural network architecture for a vision backbone includes hybrid stages with at least one state space model (SSM)-based block preceding at least one transformer block. In at least one embodiment, an SSM-based block includes parallel branches, one including an SSM and one without an SSM, and a concatenation layer for concatenating the output of each branch. In at least one embodiment, the SSM performs a parallel selective scan operation to efficiently map tokens of an input sequence to tokens of an output sequence via GPU acceleration.

A system for recovering information lost during genomic data compression employs a quality-driven approach using neural networks. The system evaluates the importance of genomic regions through a quality analysis engine that assigns quality scores, while a rate control engine determines optimal compression rates based on these scores. A specialized neural network recovers lost information from correlated genomic datasets that have undergone lossy compression, utilizing recurrent layers for feature extraction and a channel-wise transformer with attention to capture complex relationships between data channels. The neural network architecture incorporates a deblocking network that combines these components to effectively reconstruct compressed data. A decoder receives and decompresses the data, then processes it through the neural network to recover information lost during compression. This adaptive system ensures critical genomic information is preserved while maximizing compression efficiency.

The invention relates to the technical field of simulation, and particularly discloses a real-time simulation method for detecting photoelectric tracking equipment, which comprises the following steps of: performing feature decoupling processing on photoelectric signals through a dual-channel adaptive neural network architecture, and constructing an equipment mathematical model by adopting a dynamic gating fusion mechanism based on a processing result; the method comprises the following steps: establishing a multi-physics field coupling simulation environment based on a heterogeneous computing architecture, introducing an equipment mathematical model, performing three-field co-evolution through light transmission modeling, electromagnetic field distribution calculation and target motion prediction, and generating a dynamic test scene containing space-time relevance; according to the method, a mode of combining mixed feature analysis and fuzzy reasoning is adopted, multi-dimensional performance indexes are extracted from simulation data, and multi-target dynamic optimization is carried out through a strategy of combining a quantum evolution algorithm and swarm intelligence optimization; by means of a digital twin platform and a hardware-in-loop interface, real-time verification and closed-loop optimization are achieved, and the response speed and tracking precision of photoelectric tracking equipment in a complex environment are greatly improved.

The invention discloses a complex exponential signal joint spectrum reconstruction and parameter estimation method and device, and relates to the field of signal processing, and the method comprises the steps: S1, constructing noise-containing complex exponential signal training data and label data; s2, constructing a dual-module neural network model comprising a super-resolution denoising reconstruction module and a parameter prediction module; s3, training the dual-module neural network model by using the training data and the annotation data to obtain a trained dual-module neural network model; and S4, performing frequency spectrum reconstruction and parameter prediction by using the trained dual-module neural network model, and performing signal post-processing on the output to obtain estimation parameters of the angular frequency, the attenuation factor, the real part amplitude and the imaginary part amplitude. According to the invention, a cascade neural network architecture of a super-resolution denoising reconstruction module and a parameter prediction module is designed, and angular frequency detection is converted into a Gaussian distribution heat map regression task; meanwhile, a sparse activation labeling mechanism is adopted, the parameter truth value is only reserved at the spectrum peak position, and the model learning complexity is remarkably reduced.

The invention provides an evolutionary neural architecture searching method and system based on similarity agent assistance, and the method comprises the steps: 1, initializing an architecture population, and selecting an architecture with the best performance as an initial reference architecture; 2, constructing a graph neural network variant as a feature extractor; 3, constructing an agent model, and training the agent model through a joint loss function; 4, the high-potential architecture is reserved according to the fitness value, authenticity performance evaluation is carried out on the high-potential architecture, and an evaluation result is added into the training set; combining the current population with the high-performance architecture predicted and screened by the proxy model, and updating the population through an environment selection strategy; and step 5, repeating the step 3 and the step 4 until the population performance converges, and finally outputting a global optimal architecture. According to the method, the high-performance neural network architecture can be quickly searched under limited computing resources, and an efficient and intelligent novel solution is provided for automatic design and optimization of the neural network.

The invention discloses a remote sensing image fishpond extraction method based on an adaptive edge enhanced neural network. The method comprises the following steps: selecting a sentinel No.2 satellite multispectral band to synthesize a false color image; constructing a fishpond region labeling data set; establishing an adaptive edge enhanced neural network architecture, wherein the architecture fuses an edge detection module capable of learning a threshold and semantic feature extraction; an edge perception double attention module and a pyramid pooling module are integrated, and the multi-scale feature representation capability is enhanced; optimizing a training process by adopting a multi-level depth supervision and prediction fusion strategy; and finally, extracting a fishpond area through the training model and generating a standardized result. According to the method, the fishpond boundary segmentation precision is remarkably improved, the adaptability to fishponds with different scales is enhanced, the detection rate of small-scale fishponds is particularly improved, the model convergence process is accelerated through the depth supervision strategy, misclassification of similar water bodies such as rivers and ditches is effectively reduced, and the accuracy of fishpond boundary segmentation is improved. The technical problems that in the prior art, edges are fuzzy, multi-scale adaptability is poor, complex background interference is sensitive, and the omission ratio of a small-scale fishpond is high are solved.

The invention provides a lightweight potential diffusion model design method and system based on evolutionary neural architecture search, and the method employs a super-network technology to train a plurality of network architectures at the same time, combines an evolutionary genetic algorithm in an architecture search process, and trains a multi-layer perceptron as a score predictor, so as to improve the performance of the model. According to the method, the optimal architecture is efficiently screened out in a large-scale search space through the assisted evolution genetic algorithm, and the resource consumption and the calculation cost needed in the traditional architecture design process are remarkably reduced. According to the method, the pre-trained teacher model is introduced in combination with the knowledge distillation technology to supervise the training of the student model, and the potential diffusion model can be lightened while the stable performance of the potential diffusion model is ensured. Through the method, an efficient neural network architecture can be automatically designed, efficient deployment of the potential diffusion model can be realized, and a brand-new and low-cost solution is provided for optimization of the diffusion model.

The invention relates to the technical field of computer-generated holographic three-dimensional display and deep learning, in particular to an unsupervised deep learning method for realizing three-dimensional holographic display. The method comprises the steps of generating a depth map corresponding to a two-dimensional image; the depth image and the two-dimensional image are spliced in the channel dimension to serve as input of a hologram encoder, and a double-U-Net cascade neural network architecture serves as the hologram encoder; angular spectrum diffraction back propagation is carried out through the generated pure phase hologram to realize three-dimensional scene discretization reconstruction, a reconstructed image with a specified depth is obtained, a depth map is uniformly quantized to obtain a plurality of binary masks, loss calculation is carried out on the reconstructed image and a target image superposed with the corresponding depth binary masks, network parameter optimization is carried out, and a target image with the depth corresponding to the target image is obtained. And when the training of the double U-Net cascade neural network architecture is converged, the training stage is ended. According to the method, high-quality three-dimensional hologram reconstruction is realized through layered angular spectrum propagation, and the method has relatively high precision and detail reduction capability.

The invention discloses a laser radar point cloud real-time three-dimensional target detection method suitable for an unmanned vehicle. The method comprises the following steps: preparing a target detection data set; model training; updating parameters; simulation data preparation; performing model reasoning and result evaluation based on simulation data; model deployment and real vehicle data preparation; performing model reasoning and data storage based on real vehicle data; uploading and marking the data; and incremental learning and parameter updating. The three-dimensional target detection model provided by the invention specifically comprises a preprocessing module; a backbone network module; a neck network module; a feature selection module; and a detection head module. According to the three-dimensional target detection model provided by the invention, the point cloud data is converted into various aerial view images, and richer spatial information is mined; a two-dimensional distance sensing sparse convolutional layer is adopted to perform feature extraction, and multi-view feature fusion is realized through a double-branch network; a multi-scale feature pyramid network and a feature selection module are combined, so that the detection precision of a long-distance target and a small target is effectively improved; by constructing an efficient neural network architecture and combining training, simulation testing and real data incremental learning strategies, the robustness, precision and adaptive capacity of a target detection model are effectively improved, and powerful perception support is provided for an unmanned driving system.

The invention relates to the field of water quality monitoring, and discloses a water quality on-line monitoring system based on full spectrum analysis, which comprises a full spectrum water quality analysis reference module, an equipment consistency correction module, a full spectrum water quality analysis derivative module and a quality monitoring module, wherein the reference module collects full-spectrum data through reference equipment to construct a water quality analysis reference model; the correction module uses the ResNet-C neural network to correct the spectral data difference of different devices; the derivative module combines the correction data and the reference model to realize water quality index prediction; the quality monitoring module calculates a quality evaluation coefficient through a multi-dimensional index and triggers early warning; the method is based on the Lambert-Beer theorem and the neural network architecture, solves the problems of high reagent consumption and poor equipment consistency in the traditional monitoring technology, and is suitable for real-time online monitoring of the surface water quality.

Disclosed are systems and methods including software processes executed by a server that detect audio-based synthetic speech (“deepfakes”). Embodiments implement a machine-learning architecture having a diffusion model that generates purified features that are fed to a deepfake detection model. The machine-learning architecture includes input layers that convert an audio signal into a Gaussian or frequency space representation (e.g., log spectrogram) to extract a set of initial features indicative of spoofing or deepfake attacks. The diffusion model identifies adversarial noise on the audio signal in the initial features and generates purified features or clean version of the input audio signal. A deepfake detector includes a neural network architecture and classifier programmed and trained to generate a deepfake detection score and classify the audio signal as genuine or fraudulent using the purified features.

The invention provides a YOLO target detection method and system based on differential architecture search. The method comprises the following steps: firstly, constructing a candidate operation set; then, based on a YOLO network architecture, a super network is designed; then, pre-training the super network by using a differentiable architecture search method, and alternately optimizing network weight and architecture parameters; and finally, obtaining the searched subnet according to the architecture parameters, and introducing a coordinate attention mechanism to further improve the model performance. Different from a manual design method, the method has the advantages that by automatically searching the neural network architecture, tests and experiences required by manual work are greatly reduced, higher search efficiency and performance are achieved, and an efficient YOLO improved model suitable for target detection can be found more quickly.

The invention relates to the field of flood forecasting, and discloses a flood forecasting method and system, a storage medium and computing equipment, and the method comprises the steps: topological modeling: taking a river catchment area surrounded by a water diversion line and a catchment area outlet hydrometric station as nodes of a graph, constructing a connection relation of edges according to the actual flow direction of a river and the connection relation of branches, and carrying out the topological modeling; therefore, topological graph structure data adaptive to a hydrological mechanism is constructed; converting month number characteristics, generating a dynamic periodic timestamp vector, splicing the dynamic periodic timestamp vector with topological graph structure data, and inputting an attention neural network for training; the attention neural network architecture comprises two independent feature channels which are used for processing historical data and forecast data respectively, the network supports a variable modular architecture, and a flood forecast result can be output according to needs after a model is trained. According to the method, the flood prediction precision and interpretability are improved, and a high-precision and interpretable space-time joint flood prediction scheme is realized.

The invention provides a generative adversarial network architecture search method and system based on hybrid convolution operation, and the method comprises the steps: firstly designing a wide search space, then constructing a super-net integrating all candidate operations, enabling the training process of the super-net to be completely differentiable through a continuous relaxation method, and enabling the training process of the super-net to be completely differentiable; therefore, efficient training is carried out by using a gradient descent method, and finally, the optimal operation on each position is selected to obtain the optimal architecture. Different from a traditional architecture search method, the method introduces a novel hybrid convolution operation into a search space, and optimizes a search process by integrating a partial channel attention mechanism. According to the generative adversarial network architecture search based on hybrid convolution operation, the long-distance information capturing capability and architecture search speed of the network are enhanced, and compared with a traditional search method, the method has higher efficiency and better performance, and the optimal neural network architecture suitable for tasks can be found more quickly.

The invention discloses a roadway anchor net cable support parameter optimization design method based on a neural network, and belongs to the technical field of crossing of mining engineering and artificial intelligence, and the method comprises the following steps: collecting related data of a roadway, and carrying out the preprocessing of the collected data; building a dynamic self-adaptive BP neural network architecture, wherein the dynamic self-adaptive BP neural network architecture comprises an input layer introducing three cross features, a hidden layer adopting a dynamic adjustment strategy and an output layer; an anchor cable model and an anchor rod model are constructed based on a dynamic adaptive BP neural network architecture, and adaptive adjustment is performed by adopting a dynamic learning rate combined with surrounding rock complexity during model training; and inputting the five key influence indexes acquired in real time on site and the three calculated cross characteristics into the trained anchor cable model and anchor rod model to generate predicted anchor cable parameters and anchor rod parameters for on-site support design. The design efficiency of the roadway support parameter scheme is improved, the consumption of resources such as manpower is reduced, and the cost is reduced.

The invention belongs to the technical field of neural network architecture, and particularly relates to a dynamic reasoning path optimization method based on neural architecture search, which comprises the following specific steps: S1, designing a neural architecture search algorithm: firstly defining a search space, then selecting a search strategy, and then setting constraint conditions; s2, constructing a dynamic reasoning path: firstly performing input data feature analysis, then performing reasoning path guidance based on a knowledge graph, and then predicting the reasoning path by using a model according to the input data features and the knowledge graph; and S3, model training and optimization: firstly carrying out joint training, and then carrying out model compression and acceleration. According to the method, through neural architecture search algorithm design and dynamic reasoning path construction, the problem of computing resource waste is effectively solved.

The invention belongs to the technical field of nuclear radiation measurement, and relates to a liquid flash TDCR multi-nuclide beta spectrum analysis method and system based on artificial intelligence. The method comprises the following steps: constructing a numerical model of a liquid flash detector spectrometer by using a Monte Carlo technology to simulate the energy spectrum response of a single nuclide under different quenching conditions; constructing a training database by adopting a parameterized data synthesis algorithm; constructing a multi-task mixed spectrum analysis neural network model; carrying out model training by utilizing the constructed database; and inputting an actual measurement spectrogram, and outputting the multi-nuclide absolute activity, the detector efficiency and the decomposition energy spectrum contribution curve in real time by using the trained neural network model. According to the method, differential distribution characteristics and quenching response curve characteristics of the beta continuous energy spectrum are creatively fused, a neural network architecture with physical mechanism constraints is constructed, and liquid flash TDCR multi-nuclide energy spectrum characteristic decoupling and accurate activity solving are achieved under the unknown quenching condition.

The invention, which relates to the technical field of hydropower station dam safety monitoring, discloses a hydropower station dam safety monitoring data acquisition and transmission system comprising a cloud twin brain module and edge neurons. The cloud twin brain module comprises a sequence neural network engine and a reflection kernel generation module, the sequence neural network engine adopts a neural network architecture with parallel and cyclic dual representation, comprises a time mixing module and a channel mixing module, and can learn a normal operation mode of the dam from historical monitoring data; and the reflection nuclear generation module compresses the reference twin model into a lightweight reflection nuclear model and issues the lightweight reflection nuclear model to the edge device. The edge neuron comprises a micro-twinborn prediction module and a hierarchical transmission control module, and the micro-twinborn prediction module predicts a theoretical expected value of a dam state in real time and calculates a reflection deviation with an actual observation value; the hierarchical transmission control module implements a three-level response strategy according to the magnitude of the reflection deviation, transmits abstract information according to an abnormal trend, and uploads an emergency abnormality in time.